ATTENTION-ENHANCED TEMPORAL MODELING FOR PROGNOSTICS OF TURBOFAN ENGINE REMAINING USEFUL LIFE

Abstract

Advances in sensor technology and automation are shifting aviation maintenance from fixed schedules to condition-based predictive maintenance (CBPM), which leverages real-time sensor data and machine learning to anticipate failures and optimize interventions. In this study, a deep learning architecture is presented, integrating BiLSTM with a multi-head self-attention module for RUL prediction, and its performance is assessed using the NASA C-MAPSS dataset. The BiLSTM captures bidirectional temporal dependencies in degradation sequences, while the attention mechanism adaptively emphasizes critical cycles and sensor signals. Pre-processing involved piecewise RUL labelling (capped at 125 cycles), cluster-based normalization, rolling statistical features, and sliding-window sequence generation. On FD004, the BiLSTM–attention model achieved an MAE of 9.45, RMSE of 15.52, and PHM score of 3853.21, outperforming the baseline LSTM (MAE 17.80, RMSE 25.14, PHM 4211.21). On FD001, the BiLSTM–attention delivered comparable accuracy, with an MAE of 11.42, RMSE of 15.05, and PHM score of 387.15, matching or exceeding baseline performance (MAE 11.23, RMSE 15.12, PHM 395.55). These findings demonstrate that integrating bidirectional sequence modelling with adaptive attention enhances predictive robustness across varying operating conditions. The proposed approach not only achieves strong generalization but also outperforms state-of-the-art benchmarks in aircraft engine Remaining Useful Life prediction, offering practical benefits for predictive maintenance through improved safety, reduced operational costs, and extended fleet availability.